This PR introduces
- New data structure to represent kernel-level (aka node-level or
op-level) tensor sharding informaiton. I consider it as the
fundamentaion of ONNX distribtued inference.
- Building blocks for distribtued kernels implementation especially
stateless implementation for communication ops.
- Implementation of DistributedMatMul and its tests.
Code structure:
- sharding.h/.cc: Function to shard and reshard tensors (calling into
NCCL).
- sharding_spec.h/.cc: Representation of how a tensor is sharded.
- distributed_matmul.h/.cc: Implementation of tensor parallel MatMul.
Inputs and outputs are sharded across devices.
- onnxruntime_test_distributed.py: distributed operator tests.
Example of specifying sharding information
```python
@onnxscript.script()
def matmul_rs_sr_rr(tensor_x: FLOAT, tensor_w: FLOAT) -> FLOAT:
# Run MatMul by sharding x along column axis and w along row axis on
# 2 GPUs.
return MICROSOFT_OPSET.DistributedMatMul(
tensor_x,
tensor_w,
device_mesh_shape=[2],
device_mesh_elements=[0, 1],
input_shard_specs=["RS[0]", "S[0]R"],
output_shard_specs=["RR"],
)
onnx_model = matmul_rs_sr_rr.to_model_proto(
input_types=[FLOAT[2, "s"], FLOAT["s", 2]],
output_types=[FLOAT[2, 2]],
)
```
In this example, the device mesh can be visualized as 1-D tensor, `[0,
1]`. The 2nd axis of `tensor_x` is sharded across `[0, 1]` (i.e., the
0-axis of the device mesh). Similarly, the 1st axis of `tensor_w` is
sharded across `[0, 1]` as well.
C++ classes to represent tensor sharding (copied from sharding_spec.h):
```cpp
class DeviceMesh {
public:
// [Device Mesh and Tensor Sharding for Tensor Parallel]
// Device mesh is a tensor of device indices.
// A tensor can then be partitioned along specific mesh axes.
//
// Assume we have 4 GPUs indexed by 0, 1, 2, and 3.
// Let's consider some examples.
// 1. 1D device mesh [0, 1, 2, 3]. In this case,
// device_mesh_shape is [4] and device_mesh_elements
// is [0, 1, 2, 3].
// If we want to shard a 2-D tensor along its axis 1, the
// corresponding sharding spec is a string "RS[0]".
// 2. 2D device mesh [[0, 1], [2, 3]]. In this case,
// device_mesh_shape is [2, 2] and device_mesh_elements
// is [0, 1, 2, 3].
// If we want to shard a 2-D tensor's
// rows along mesh axis 1 and
// columns along mesh axis 0, the
// corresponding sharding spec is a string "S[1]S[0]".
// If that 2-D tensor's value is np.array([[5, 6], [7, 8]]),
// GPU 0/1/2/3 owns 5/7/6/8. Below is a visualization the sharding
// proccess.
// - Start with a 2-D device mesh [[0, 1], [2, 3]] and
// a 2-D tensor [[5, 6], [7, 8]]
// - GPU: [[0, 1], [2, 3]], Tensor: [[5, 6], [7, 8]]
// - Split GPU mesh along axis 1 and tensor along
// axis 0 for "S[1]" in "S[1]S[0]"
// - GPU: [[0], [2]], Tensor: [[5, 6]]
// GPU: [[1], [3]], Tensor: [[7, 8]]
// - Split GPU mesh along axis 0 and tensor along
// axis 1 for "S[0]" in "S[1]S[0]"
// - GPU: [[0]], Tensor: [[5]]
// - GPU: [[2]], Tensor: [[6]]
// - GPU: [[1]], Tensor: [[7]]
// - GPU: [[3]], Tensor: [[8]]
// Actual shape of device mesh represented by `device_mesh_elements`.
std::vector<int64_t> device_mesh_shape;
// Flattened device mesh.
std::vector<int64_t> device_mesh_elements;
};
class AxisPartitionSpec {
// [Device Mesh and Tensor Sharding for Tensor Parallel]
// This class is the in-memory representation of
// 1. if a tensor is sharded or not (aka replica), and
// 2. which tensor axis is shard by which device mesh axis.
// Let's consider sharding 2-D tensor along column axis on
// device mesh [0, 1] as an example.
// The required sharding spec RS[0] can be represented by
// - AxisPartitionSpec(Condition::Replica, -1)
// - AxisPartitionSpec(Condition::Shard, 0)
public:
// Status of a tensor axis.
// A tensor axis can be either sharded or replicated
// along a device mesh axis.
enum class Condition { Replica,
Shard };
// This field tells if a tensor axis is sharded or not.
Condition cond;
// If a tensor axis is sharded, this field tells which device
// mesh axis to distribute the shards along.
// If a tensor axis is not sharded, this field is ignored.
int device_mesh_axis;
// A helper to construct a replica spec for a tensor axis.
static AxisPartitionSpec CreateReplica() {
return AxisPartitionSpec(Condition::Replica, -1);
}
// A helper to construct a sharding spec for a tensor axis.
// This tensor axis is sharded along `device_mesh_axis` in device mesh.
static AxisPartitionSpec CreateShard(int device_mesh_axis) {
return AxisPartitionSpec(Condition::Shard, device_mesh_axis);
}
};
class TensorPartitionSpec {
// [Device Mesh and Tensor Sharding for Tensor Parallel]
// TensorPartitionSpec holds a collection of AxisPartitionSpec and an
// associated DeviceMesh. It is responsible for determining how a tensor
// should be partitioned across a device mesh.
//
// Example 1: RS[0]
// In this scenario, `axis_specs` would contain two `AxisPartitionSpec` objects.
// - The first object is a Replica, denoting that the first axis of the tensor is
// not sharded but is instead replicated.
// - The second object is a Shard along the 0-th axis of the device mesh. It denotes
// that the second axis of the tensor is sharded along the first axis of the
// device mesh.
//
// Example 2: S[0]RR
// In this scenario, `axis_specs` would contain three `AxisPartitionSpec` objects.
// - The first object is a Shard along the 0-th axis of the device mesh, indicating
// that the first axis of the tensor is sharded along the first axis of the
// device mesh.
// - The second and third objects are Replicas, indicating that the second and third
// axes of the tensor are not sharded but are instead replicated.
public:
// axis_specs[i]: AxisPartitionSpec for tensor axis i. For a 2-D tensor,
// axis_specs[0] is for row axis and axis_specs[1] is for
// column axis. axis_specs[i].device_mesh_axis = j means that
// tensor axis i is sharded along device mesh axis j.
std::vector<AxisPartitionSpec> axis_specs;
// device_mesh: DeviceMesh for sharding the associated tensor.
// Read [Device Mesh and Tensor Sharding for Tensor Parallel] in DeviceMesh's comment.
DeviceMesh device_mesh;
};
```
### Description
<!-- Describe your changes. -->
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
### Description
Introduce collective ops into onnxruntime inference build, including
1) AllReduce and AllGather schema in contrib op, controlled by USE_MPI
flag
2) AllReduce and AllGather kernel in cuda EP, controlled by ORT_USE_NCCL
flag
### Motivation and Context
Enable the collective ops in onnxruntime inference build so we have the
ability to run distributed inference with multiple GPUs.
The original ncclAllReduce ops in training build require quite complex
configurations, which is not suitable for inference case, and it already
broken. so we introduce a new implementation.
---------
Co-authored-by: Cheng Tang <chenta@microsoft.com@orttrainingdev9.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
### Description
1. Renames all references of on device training to training apis. This
is to keep the naming general. Nothing really prevents us from using the
same apis on servers\non-edge devices.
2. Update ENABLE_TRAINING option: With this PR when this option is
enabled, training apis and torch interop is also enabled.
3. Refactoring for onnxruntime_ENABLE_TRAINING_TORCH_INTEROP option:
- Removed user facing option
- Setting onnxruntime_ENABLE_TRAINING_TORCH_INTEROP to ON when
onnxruntime_ENABLE_TRAINING is ON as we always build with torch interop.
Once this PR is merged when --enable_training is selected we will do a
"FULL Build" for training (with all the training entry points and
features).
Training entry points include:
1. ORTModule
2. Training APIs
Features include:
1. ATen Fallback
2. All Training OPs includes communication and collectives
3. Strided Tensor Support
4. Python Op (torch interop)
5. ONNXBlock (Front end tools for training artifacts prep when using
trianing apis)
### Motivation and Context
Intention is to simply the options for building training enabled builds.
This is part of the larger work item to create dedicated build for
learning on the edge scenarios with just training apis enabled.
* upgrade cuda version on ci pipelines
* keeping folder name same
* keeping folder name same
* setting manual seed for primitive test case
* resolving comments
* changing atol and rtrol only for test case
Co-authored-by: Adam Louly <adamlouly@microsoft.com@orttrainingdev7.d32nl1ml4oruzj4qz3bqlggovf.px.internal.cloudapp.net>
* fix build - python.h not found
* disable --build_shared_lib for ortmodule tests
* fix
* fix the build flag
* disable --build_shared_lib for training path (not only for ortmodule)
* fix missing test model files
* disable test CApiTest.test_custom_op_library when ENABLE_TRAINING_TORCH_INTEROP is ON
* enable custom_op_library build
* fix build
* fix
* merge master and fix build failure
* build onnx_test_runner when onnxruntime_ENABLE_TRAINING_TORCH_INTEROP is ON
* resolve comments
* use --enable_training_torch_interop to replace "onnxruntime_ENABLE_TRAINING_TORCH_INTEROP=ON"
ORTModule requires two PyTorch CPP extensions that are currently JIT compiled. The runtime compilation can cause issues in some environments without all build requirements or in environments with multiple instances of ORTModule running in parallel
This PR creates a custom command to compile such extensions that must be manually executed before ORTModule is executed for the first time. When users try to use ORTModule before the extensions are compiled, an error with instructions are raised
PyTorch CPP Extensions for ORTModule can be compiled by running:
python -m onnxruntime.training.ortmodule.torch_cpp_extensions.install
Full build environment is needed for this
1. Remove some unused code and simplify tools/ci_build/github/linux/run_dockerbuild.sh.
2. Enable Nuget CUDA tests. The original design was we could leverage Directory.Build.props and let cmake generate the required properties(USE_CUDA/...) there. However, in nuget packaging pipeline we test the package on a different host that doesn't run cmake command and doesn't have the auto-generated Directory.Build.props file.
* clean up builds for interop_torch
* add python dependency for executables
* disable onnxruntime_ENABLE_TRAINING_TORCH_INTEROP by default; enable it in ortmodule GPU training pipeline only
* disable training unrelated tests when torch interop is enabled
* simplify the python dependency.
* clean up and fix
1. Fix training e2e pipeline. The failure was caused by my recent change #7632. The fix is adding "--cmake_extra_defines CMAKE_CUDA_ARCHITECTURES=70" to the build parameters because the machines are with V100 GPUs.
2. Simplify Nuphar pipeline. It doesn't need to install a separated ONNX version(1.5.0)
3. Fix a problem that run_dockerbuild.sh ignored OS version parameter. Now because it starts to take effect, I also set python version to the system default one(3.8 for ubuntu 20.04)
1. Update manylinux build scripts. This will add [PEP600](https://www.python.org/dev/peps/pep-0600/)(manylinux2 tags) support. numpy has adopted this new feature, we should do the same. The old build script files were copied from https://github.com/pypa/manylinux, but they has been deleted and replaced in the upstream repo. The manylinux repo doesn't have a manylinux2014 branch anymore. So I'm removing the obsolete code, sync the files with the latest master.
2. Update GPU CUDA version from 11.0 to 11.1(after a discussion with PMs).
3. Delete tools/ci_build/github/linux/docker/Dockerfile.manylinux2014_cuda10_2. (Merged the content to tools/ci_build/github/linux/docker/Dockerfile.manylinux2014_cuda11)
4. Modernize the cmake code of how to locate python devel files. It was suggested in https://github.com/onnx/onnx/pull/1631 .
5. Remove `onnxruntime_MSVC_STATIC_RUNTIME` and `onnxruntime_GCC_STATIC_CPP_RUNTIME` build options. Now cmake has builtin support for it. Starting from cmake 3.15, we can use `CMAKE_MSVC_RUNTIME_LIBRARY` cmake variable to choose which MSVC runtime library we want to use.
6. Update Ubuntu docker images that used in our CI build from Ubuntu 18.04 to Ubuntu 20.04.
7. Update GCC version in CUDA 11.1 pipelines from 8.x to 9.3.1
8. Split Linux GPU CI pipeline to two jobs: build the code on a CPU machine then run the tests on another GPU machines. In the past we didn't test our python packages. We only tested the pre-packed files. So we didn't catch the rpath issue in CI build.
9. Add a CentOS machine pool and test our Linux GPU build on real CentOS machines.
10. Rework ARM64 Linux GPU python packaging pipeline. Previously it uses cross-compiling therefore we must static link to C Runtime. But now have pluggable EP API and it doesn't support static link. So I changed to use qemu emulation instead. Now the build is 10x slower than before. But it is more extensible.
* Install and use conda on ortmodule CI pipelines
* Update build script to install onnxruntime wheel before running unit tests
* Remove python 3.5 from install_python_deps
* Pinning deepspeed version to 0.3.15
